feat: auto-inject builtin agent skills and CLAUDE.md at startup

[NathanFlurry]

Summary

• Adds a builtin-skills system that auto-injects process management knowledge into agents at startup
• Embeds SKILL.md (API docs) and CLAUDE.md (behavioral directives) at compile time via include_str!
• Writes skills to ~/.claude/skills/ and injects a managed section into ~/.claude/CLAUDE.md with idempotent marker-based updates
• Teaches agents to use the process HTTP API for long-lived processes (dev servers, watchers, Flask apps) while using Bash directly for one-shot commands (builds, tests)
• Substitutes $SANDBOX_AGENT_URL with the actual server URL at write time
• Disableable via SANDBOX_AGENT_SKIP_BUILTIN_SKILLS=1

Also includes workspace recovery: factory/ removal, font updates to Manrope, and pnpm-lock cleanup.

Key files

• server/packages/sandbox-agent/builtin-skills/CLAUDE.md — behavioral directives injected into agent CLAUDE.md
• server/packages/sandbox-agent/builtin-skills/sandbox-agent-processes/SKILL.md — process API documentation skill
• server/packages/sandbox-agent/src/builtin_skills.rs — compile-time embedding and write logic
• CLAUDE.md — added rules for keeping skills in sync with API changes

Test plan

• Build and start server, verify ~/.claude/skills/sandbox-agent-processes/SKILL.md is written
• Verify ~/.claude/CLAUDE.md has the SANDBOX-AGENT-MANAGED section with correct URL
• Verify SANDBOX_AGENT_SKIP_BUILTIN_SKILLS=1 prevents injection
• Test agent uses process API for dev servers / watchers (long-lived)
• Test agent uses Bash directly for builds / tests (one-shot)

🤖 Generated with Claude Code

[railway-app]

🚅 Deployed to the sandbox-agent-pr-223 environment in sandbox-agent

Service	Status	Web	Updated (UTC)
website	😴 Sleeping (View Logs)	Web	Mar 12, 2026 at 10:26 pm